Modular Dynamic Neural Network: A Continual Learning Architecture

Learning to recognize a new object after having learned other objects may be simple task for human, but not machines. The present go-to approaches teaching machine set of are based on the use deep neural networks (DNN). So, intuitively, solution fly should DNN. problem is that trained DNN weights used classify initial extremely fragile, meaning any change those can severely damage capacity perform recognitions; this phenomenon known as catastrophic forgetting (CF). This paper presents (DNN) continual learning (CL) architecture deal with CF, modular dynamic network (MDNN). presented consists two main components: (a) ResNet50-based feature extraction component backbone; and (b) classification component, which multiple sub-networks progressively builds itself up in tree-like structure rearranges it learns over time such way each sub-network function independently. contribution strongly its training feature. allows classes added while only altering specific previously forgotten. Tests CORe50 dataset showed results above state art CL architectures.